My Own Private Nightlife: Understanding Youth PersonalSpaces from Crowdsourced Video
THANH-TRUNG PHAN∗, Idiap Research Institute and EPFL, SwitzerlandFLORIAN LABHART, Idiap Research Institute, SwitzerlandDANIEL GATICA-PEREZ, Idiap Research Institute and EPFL, Switzerland
Private nightlife environments of young people are likely characterized by their physical attributes, particularambiance, and activities, but relatively little is known about it from social media studies. For instance, recentwork has documented ambiance and physical characteristics of homes using pictures from Airbnb, butquestions remain on whether this kind of curated data reliably represents everyday life situations. To describethe physical and ambiance features of homes of youth using manual annotations and machine-extractedfeatures, we used a unique dataset of 301 crowdsourced videos of home environments recorded in-situ byyoung people on weekend nights. Agreement among five independent annotators was high for most studiedvariables. Results of the annotation task revealed various patterns of youth home spaces, such as the typeof room attended (e.g., living room and bedroom), the number and gender of friends present, and the typeof ongoing activities (e.g., watching TV alone; or drinking, chatting and eating in the presence of others.)Then, object and scene visual features of places, extracted via deep learning, were found to correlate withambiances, while sound features did not. Finally, the results of a regression task for inferring ambiances fromthose features showed that six of the ambiance categories can be inferred with R2 in the [0.21, 0.69] range.Our work is novel with regard to the type of data (crowdsourced videos of real homes of young people) andthe analytical design (combined use of manual annotation and deep learning to identify relevant cues), andcontributes to the understanding of home environments represented through digital media.CCS Concepts: • Human-centered computing → Ubiquitous and mobile computing; Ubiquitous andmobile computing design and evaluation methods.
Additional Key Words and Phrases: Youth; Mobile crowdsensing; Ambiance; Nightlife; HomeACM Reference Format:Thanh-Trung Phan, Florian Labhart, and Daniel Gatica-Perez. 2019. My Own Private Nightlife: UnderstandingYouth Personal Spaces from Crowdsourced Video. Proc. ACM Hum.-Comput. Interact. 3, CSCW, Article 189(November 2019), 34 pages. https://doi.org/10.1145/3359291
1 INTRODUCTIONThe home environment is an important subject of study in several social sciences including psy-chology and geography, as well as architecture and design, and more recently computing [4], [5].Private spaces at home include common living spaces in households (living room and kitchen), but
∗This is the corresponding author.
Authors’ addresses: Thanh-Trung Phan, Idiap Research Institute and EPFL, Switzerland, [email protected]; Florian Labhart,Idiap Research Institute, Switzerland, [email protected]; Daniel Gatica-Perez, Idiap Research Institute and EPFL, Switzerland,[email protected].
also individual personal rooms (bedrooms). It is known that young people appropriate their privatespaces and manifest aspects of their personal beliefs and traits in this way [52], [53], [54].One important feature of place (private and otherwise) is ambiance. This is defined as “the
character of a place or the quality it seems to have” [24] and used for both indoor environments[75], [68], [60], [16], [12] and outdoor environments [62], [65]. In the context of commercialspaces, ambiance plays an important role in customer behavior related to shopping [45], foodchoices [10], [82] or hotel experiences [21]. Regarding the ambiance of personal spaces, previouswork has shown that ambiance mediates other factors, like gender [69] and personality [30], onchoices made on physical and environmental characteristics. Understanding home spaces is arelevant domain that has not been fully studied in social computing. A better understanding ofphysical and social attributes and ambiance of personal spaces could have various implicationsfor social computing research as a part of an agenda on living spaces and well-being. For example,homes can be reconfigured by their inhabitants with respect to decoration, spatial organization offurniture, light, and music, thus inducing more appropriate ambiances for certain activities andsocial interactions at home (e.g. a romantic dinner vs. an end-of-year party). Designing systems thatboth recognize physical and social attributes and support users to reconfigure their home spacesbased on their specific goals is a relevant application. This could integrate the many perspectivesexisting in psychology, architecture, human geography, and public health, with the availabilityof environmental and mobile sensors and social media, and is a particularly interesting angle tounderstand and support youth practices. Furthermore, many traditional studies have collectedinformation of personal spaces by using paper-and-pencil questionnaires and interviews. Thepotential of collecting in-situ information of home spaces (physical and social attributes of theenvironment) through technological means could add to the existing set of research tools.Recent work on recognition of indoor ambiance [75], [60] has used still images from online
social systems like Foursquare [75] or Airbnb [60]. Social psychologists have also investigatedimpression formation on home environments [32], [30]. Yet, gaps in the existing body of workemerge as most previous work has been conducted using social media data that either (1) mightlack diversity in the representations of private residences [64], [22], as they are naturally focusedon outdoor and commercial spaces; or (2) might be beautified, e.g. on Airbnb and similar sites, dueto the intrinsic motivations to create and share such images [60]. To investigate home ambiancein a naturalistic setting, a different research direction could use crowdsourcing to collect in-situvideos of the personal environments inhabited by volunteers, which will yield vivid image andsound information, while reducing certain motivations of the video makers (e.g. performative orcommercial) that could otherwise affect the generated content. Compared to previous work, ourwork uses crowdsourced video data of young people in their personal spaces during weekendnights. This provides a new view of youth nightlife activities in the private sphere, which enrichesthe kind of information that traditional methods in the social sciences can provide, in terms oftemporal and scene granularity as well as scale.The overall aim of the present work is to understand the characteristics of private spaces in
youth nightlife in the weekend by investigating physical environment features and ambiances ofhome. Specifically, using a combination of human annotation and machine learning (computervision & audio processing), we address the following research questions:
RQ1: Given crowdsourced videos recorded at home spaces by young people at night, whatpatterns of physical and ambiance attributes of youth home spaces can be revealed by manualcoding of videos using external annotators and machine-extracted features?
RQ2: Given machine-extracted features of videos at youth home spaces, can these features inferthe perceived ambiance of such spaces?
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:3
To address these questions, we use a crowdsourced dataset about nightlife, involving the partici-pation of 241 young people, aged 16 - 25 years old, in the two main Swiss hubs for nightlife (Zurichand Lausanne) [72]. To obtain detailed insights on the locations attended and their related ambiance,participants were requested to record panoramic 10-second video clips of their environment at thestart of the night and whenever they changed locations. In total, 841 videos were collected on 10weekend nights. In this dataset, a significant portion of locations documented were private places[72], which provides a unique snapshot of how weekend nights are experienced by youth in theirprivate environments. We design and implement an annotation task by asking external annotatorsto watch video clips. To build up this questionnaire, we have adopted several dimensions fromrelated work [29], [31], [74], [60]. As a result, we generated a labeled dataset of 301 video clips inpersonal environments which contain richer in-situ information than what is often captured inquestionnaires or surveys used in previous work, and manually annotated attributes of privatespaces for our analysis.
Our paper has the following contributions:(1) To address RQ1, we use a 301-video dataset of home spaces collected by Swiss young people
on weekend nights. Our dataset contains video and audio files. A set of five independent ratersannotated all videos with a rich set of questions, including physical attributes, social attributes, andambiance. The results show that the video dataset can be consistently assessed by external raters,with at least moderate agreement, and in many cases with good or excellent agreement. Detailedanalyses of the annotations produce several relevant results. First, we show that activities likeeating, drinking, and entertainment (chatting, watching TV, and using digital portable devices) areall popular among young people, but with fluctuations over the night period. Second, we found asubstantial number of cases where young people are alone and where home place loudness (chatterand music) is low. For those cases in which people socialize, we observed a same-sex trend betweenstudy participants and their companions. Third, we performed a correlation analysis among theambiance attributes that showed two main opposite dimensions, namely places perceived as large,colorful, comfortable, festive, stylish, and unique; and a second category of places perceived asconfined, simple, and boring. Dark and bright ambiances did not show significant correlation withthe rest of the ambiance attributes. Finally, we use deep learning models applied on the audio andvideo tracks to extract automatic features to represent private spaces at the level of objects, scenes,and sounds; the results indicate the feasibility of using deep learning to produce generic semanticdescriptions of home environments, although in several cases interpretation remains an issue.
(2) To address RQ2, we find that several of the 1000-object, 365-scene, and 527-sound classes usedin our work have a particular correlation with specific ambiances. Finally, we use a machine learningpipeline to automatically infer ambiances of private spaces (as a regression task) using featuresinformative of sounds, objects, and scenes. The results show that object and scene classes can predictsix ambiances with R2 between 0.21 and 0.69: space capacity (large/spacious vs. cramped/confined),brightness (bright/well-lit vs. dark/badly-lit), comfortable/cozy, and dull/simple.
The paper is organized as follows. Section 2 discusses related work. Section 3 presents the datacollection and annotation process. Section 4 presents the in-depth analysis of private spaces basedon the manual annotations. Section 5 presents the approach based on deep learning to extractvisual and sound features of videos, examines the correlation between ambiance and the previouslyextracted cues, and presents experiments on automatic ambiance inference. Section 6 discusses thefindings and limitations of our work from the perspectives of social computing. Section 7 concludesthe paper.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:4Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
2 RELATEDWORKOur work is related to a body of work from various disciplines examining issues of urban nightlifeand youth; characterization of private spaces; and ambiance modeling. Each of these themes isdiscussed in the next subsections.
2.1 Urban nightlife and youthWork in geography has studied the urban night period, often with qualitative methods [90]. Theauthors in [84] , [33], [14] also studied the dynamics surrounding youth experiences and urbannightlife. There is other work that has investigated the phenomena of human mobility and spaceusage in urban areas [59], [18], [81], [37]. From the perspective of alcohol consumption and urbanyouth, researchers investigated pubs and bars [25], house parties [40] and public spaces [23].Especially, [90], [35], [11], [95] studied alcohol consumption from “pre-loading” (drinking beforegoing out for the night) to excessive drinking with risky consequences. In contrast with theseworks, our paper aims to understand the characteristics and activities of the nightlife of youth intheir home environments based on captured videos of the private spaces, contributed by the studyparticipants in a crowdsensing setting.
2.2 Place characterization and private spacesRegarding place characterization, the authors of [92] used mobile sensors, i.e. audio signals toinfer occupancy, human chatter, music, and noise of places. Meanwhile, the authors of [20] aimto categorize places by using audio signals and images. Chon et al. [19] collected 48,000 placevisits from 85 participants in Seoul to study the coverage and scaling properties of place-centriccrowdsensing.As a private space, the home is an environment where many social activities of young people
unfold [4], [5]. In geography, Abbott et al. [4] investigated perceptions of young people about homeas an idealized social construct and as a private space. Abbott et al. later investigated the socialconstructs of ‘home’ and ‘neighborhood’ as private and public spaces, in the context of leisureactivities performed by young adolescents [5]. These studies used standard methods based onrecall-based surveys. From a technical perspective, work in ubiquitous computing has developedapproaches for place characterization, which use mobile sensors like microphones to extract audiosignals through which certain features like human chatter and music can be inferred [92], ora combination of audio signals and still images that capture snapshots of everyday places [19],[20]. This body of work, however, has been largely focused on understanding outdoor spaces,often with goals of automatic place recommendation for urban users. In contrast to this work,we investigate how attributes of the home environment of young people are depicted on videosrecording snapshots of weekend nights (a period of intense socialization among youth [4], [5])using both human observers and machine-generated descriptors of the home environments.
2.3 Home spaces and activitiesHome is conceived in different ways, including a physical space (house/apartment), someone’splace of origin, or the place where people feel they belong [3]. Regarding the place of origin orwhere a person feels as belonging to, home is a site of ‘shelter’ [34] or a ‘meaningful’ place withmultiple experiences through which people feel belonging [6], [87]. Home can not only be a fixedspace but also an urban area, e.g., a street in town or a popular area in the city [6]. Home can alsobe a material place where young people live with their family [13], [77], or a student home ordormitory where students study or live away from their parents [36]. In our research, we aim to
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:5
understand home as a personal space where young people spend their time alone or with friendson weekend nights.Home is one of the places where youth spend their leisure time e.g., watching TV, listening to
music [77], playing physical games [71], or drinking before going out with their friends at publicplaces in the city [95]. Many people also socialize at their friends’ or family’s house [35], whichemphasizes the importance to understand these practices, as the use of rooms and spaces at homecan be influenced by architectural constraints, culture, an individual’s daily life [7], or even mentaldistress [88]. Baillie et al. [9] studied four spaces in the home, including communication, work,leisure (private) and leisure (public) along with their utility to people living there. In our work, weinvestigate multiple dimensions of home of youth on weekend nights, including physical attributes(e.g. room types, brightness, music), social attributes (people present in the home environment),and ambiance (e.g. festive or fun).
2.4 Ambiance in architecture and psychologyThe roles of interior architecture and design on human behavior have been studied in severaldisciplines, and provide background about the way humans interact with their living spaces. Thecharacteristics of the places where we live, including space quality, interior design, and colors,affect how we feel, and reflect personal and social constructs [28]. Three main factors discussed by[28] influence living spaces: identity claims, thought/feeling regulators, and behavioral residues.Interior ambiance, i.e., "the character of a place or the quality it seems to have" [24] can havespecific effects on people’s behavior.In the context of personal spaces, physical and environmental cues can reveal characteristics
related to gender [69] and personality [30]. A common method used in psychology [69], [30]involves asking observers to manually rate physical spaces, which is an approach applicable tosmall-scale studies. Our work uses this methodology, and expands it by using automatic analysis tocharacterize the content of videos using state-of-art deep learning methods.
Ambiance has also been studied in public and commercial spaces. Quercia et al. [66] presented acrowdsourcing project related to ambiance-related constructs in the outdoor space, which studiedhow visual cues, color, and texture have effects on London neighborhoods along three perspectives:beautiful, quiet, and happy. In [45], physical and decoration cues had effects on the shoppingbehaviors of customers, because people’s emotions and behaviors can be affected by these places’ambiance. Ambiance cues like color, brightness, and style have an important impact on customeremotions at hotels [21], or on food intake and food choice at restaurants [82]. For instance, [10]showed that decorating the ambiance of a pasta restaurant with a distinctive Italian feeling canmake customers order more food.Specifically for home environments, the Personal Living Space Cue Inventory (PLSCI) [29]
describes personal living spaces, including 42 physical attributes and the ambiances of the spacealong with a checklist of 100 individual items. PLSCI is used by [17], [30] to study various questionsin environmental psychology. We also adopt PLSCI for designing the video ambiance questionnairefor our study about home spaces of youth, which is discussed in Section 4.
2.5 Indoor ambiance inference in social computingSeveral works have proposed methods to automatically recognize indoor ambiance from socialmedia data. By observing the avatar pictures of Foursquare users, the work in [31] showed thatpeople can identify place ambiance, clientele, and their activities with some degree of reliability.The work in [68] also used 4sq profile pictures to infer place ambiance using aesthetics, colors,emotions, demographics (age and gender), and self-presentation. Although the number of datapoints used in this work was small (N=49), it showed promise for place ambiance inference. Using
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:6Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
data from Foursquare, the work in [74] generated crowdsourced annotations on an image corpus tostudy 13 ambiance dimensions. This dataset was later used to apply traditional visual features (color,GIST, HOG) and features extracted from a pre-trained CNN for ambiance inference [75]. The workin [12] further examined the problem of ambiance recognition through scene semantics, assumingthat there are visual cues within scenes that can be extracted using a scene-centric semantic parser.We also adopt this assumption in our paper for conducting annotation on ambiance by asking ratersto watch videos. However, the datasets used in previous work are images from Foursquare places,thus covering restaurants, bars, cafes, etc.. In this paper, we work with substantially different data,namely with videos capturing private spaces during 10 seconds, and through the combination ofmanual annotation results of ambiance based on the observation of the captured videos, and onsemantic video cues extracted from deep learning models.Airbnb is a social platform for hospitality that shows home environments to possible guests
through photos. Ikkala et al. [38] conducted a qualitative research of hospitality exchanges onAirbnb. The study found that hosts on Airbnb have both financial and social reasons. In detail,money plays a role in supporting hosts in their efforts to manage social interaction, select guestsconsistent with their preferences, and control the volume and type of demand of visitors. In whatconstitutes the closest work to ours, [60] used a dataset of 1200 Airbnb venues represented by threeimages of each place to infer ambiances from visual features extracted from deep learning models.This work is an inspiration to our work, with one fundamental difference, namely that the visualdata responds to very different motivations: crowdsensing for scientific research in our case, andillustrating home places for monetary purposes on Airbnb. This translates into rather differentvisual content: on Airbnb, images are curated to appear as appealing as possible to viewers; inour work, the videos produced by youth on weekend nights are unfiltered (except for reasons ofsensitive situations and privacy) and non-beautified (as the study participants are sharing this datafor research only and not for performative purposes as is often the case on Instagram and othersocial media).To the best of our knowledge, our work extends the current understanding of private nightlife
settings with respect to physical attributes at homes, activities of young people, and ambiances,building upon previous work in the CSCW, social computing, and ubicomp literature [77], [71],[67], [9], [66], [38]. In Table 1, we summarize the most closely related work and distinguish whatwe contribute to this domain.
3 DATA COLLECTIONOur work uses data from the Youth@Night project [72], which aimed at studying young peoplenightlife behavior in Switzerland using a smartphone application [72], [48]. This section providesan overview of the study design, the data collection procedure, and the specific data we use in thiswork.
3.1 Study design3.1.1 Study context. Participants were recruited in Zurich and Lausanne, two of the four largestSwiss cities [72],[47] and the two main hubs of nightlife activities [58], [61]. They were approachedby small groups of research assistants on the street between 8 PM and midnight in September2014. In order to obtain a representative sample of nightlife goers, participants were recruited inpopular areas (e.g., nightlife districts, public parks, streets), pro-rata of the area popularity at thecity level. Quotas of people to recruit per area were determined using geo-localized venue datafrom Foursquare [73], and were validated with local experts (social workers and police). Eligibilitycriteria for participation were being aged between 16 and 25, owning an Android phone, havingbeen out in the city at least once in the past month, and have consumed alcohol at least once in the
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:7
Work Goal Data Tasks Finding
[5](2001)
Observe young peo-ple’s favourite placesand associated leisureactivities at home andneighbourhood.
Young people wanttheir homes to befriendly, spacious,modern, and quiet, theyhangout with friendsat home and friends’homes.
[92](2014)
Infer the ambiance ofbusiness places fromaudio recordings
150 audio traces of in-door business and exter-nal surveys
Regression task forinferring the levelof occupancy, hu-man chatter, music,and noise levels us-ing audio features
Classification perfor-mance of ambiance at79% accuracy
[68](2015)
Determine which vi-sual cues of profilepictures can predictplaces’ ambiances
Ambiance surveys of 49places, with 250 annota-tions on 25 profile pic-tures on each place.
Regression task forpredicting place am-biance using profilepictures’ features
Predict ambiance basedon faces at 78%
[74](2015)
Investigate whichtypes of social mediaimages best conveyindoor ambiance
50K images from 300places on Foursquare,and 13 ambiance labels.
Interannotatoragreement (ICC)analysis ans corre-lation analysis.
All 13 dimensions haveICC>0.5
[75](2016)
Infer impressions ofplace ambiance, us-ing generic and deeplearning features
45,000 Foursquare im-ages from 300 popularplaces in six cities
Regression task forinferring ambianceusing machined-extracted features
Inferring place am-biance is feasible with amaximum R2 of 0.53
[12](2017)
Examine correlationof visual cues with am-biance of Foursquareimages to automati-cally infer place am-biance
50K Foursquare imagesand 20K scene centricimage dataset
Regression task forinferring ambianceusing deep learningfeatures
Ten of the ambiancescan be inferred usingscene objects and demo-graphic attributes
[60](2018)
Predict ambiancefrom pictures oflistings on Airbnb
1200 Airbnb listingsand crowdsourcedannotations of images
Regression task forinferring ambianceusing deep learningfeatures
Ambiance can be in-ferred with R2 up to0.42
Ourwork
Describe youthpersonal spaces bymeans of crowd-sourced videosrecorded in-situ.Labels are differentthan all above workexcept [60].Infer ambiance atyouth personalspaces from physicalattributes
301 videos recordedin participants’ homespace on weekendnights.Manual annotationsof the 301 videos by 5independent annota-tors and CNN-basedextraction of visual andaudio descriptors.
Descriptive and cor-relation analysesof ambiances andphysical features ofhome spaces.Regression taskto infer ambianceusing machined-extracted features
Living room, bedroom, kitchen, anddining room are allrepresented at homeon weekend nights.Top activities includedrinking, chatting,watching TV, andeating. Home ambiancewas often describedas quiet and simplydecorated. Regressionfor ambiance inferenceachieved R2 between0.21 − 0.69.
Table 1. Comparison between previous work and our work.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:8Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
past month (legal drinking in Switzerland, as in many other European countries, is 16 for beer andwine). The study protocol was approved by the ethical review boards of Vaud and Zurich cantons,and authorization to recruit on the street was obtained from the local authorities.
3.1.2 Data collection. The study took place on Friday and Saturday nights between Septemberand December 2014. Participants were required to download and install Youth@Night applications.The survey logger application allowed participants to document, in real time, various aspects oftheir night, such as the locations attended (e.g. home, park, bar/pub), the type of drinks consumed,and 10-second video clips of their environment from 8 PM to 4 AM. Meanwhile, the sensor loggerapplication, a background running app without any user interaction, collected many types ofsensors and log data, such as GPS coordinates, accelerometer, and battery status [72],[47]. In thiswork, we will only use data from the survey logger application.
Questionnaires and sensor datasets were automatically uploaded to a back-end server whenparticipants’ smartphones had access to Wifi. Whenever the data was successfully uploaded, it wasremoved from the device. The participants could choose to manually upload data in case there wasa problem with the automatic upload. At the end of the study, participants were paid 100 CHF ifthey documented at least 10 weekend nights. Participants completing less than 10 evenings with aminimum of three nights were paid on a pro-rata basis.
After the app-based data collection fieldwork, 40 qualitative interviewswere conductedwith studyparticipants and focused on their experiences with the smartphone application, their experiencesof nights out, and the ways in which mobile technologies shape contemporary nightlife [85], [86].
3.2 Measures3.2.1 Video clips of environments. The survey logger application contained different questionnairesand media to capture participants nightlife behaviors, the locations attended, and the characteristicsof their surrounding environment (see [48] for an overview of the different kinds and sequencesof questionnaires). Participants were instructed to document any weekend night, including thoseduring which they did not drink or did not go out, in order to have an overall representation of thedifferent activities and events taking place on weekends. In the present work, we use the short videoclips collected with the application at specific times of the night: whenever participants had theirfirst drink (alcoholic or non-alcoholic) after 8 PM, and whenever they had a new drink (alcoholic ornon-alcoholic) in a new location, they were required to indicate the type of location they attended(e.g. bar/pub, parks, home) and to record a 10-second video clip, which captured a panoramaof their environment by slowly turning from left to right in landscape format. Participants thusrecorded videos in varied environments, including pubs, clubs, public parks, means of transportation,and homes [72]. In case they were not able to record video (e.g., forbidden, felt uncomfortable),participants were told to skip the task and specify the reasons for it. Overall, participants recordedvideos in 68% for the cases, while reasons for not recording were mostly because they did not itfeel it as appropriate or safe [48]. Each video file was stamped with its time of submission. In total,843 videos were collected from 204 participants on 646 participant-nights.
3.2.2 Annotation of home environments. After the fieldwork, we designed an annotation task to getqualitative information on the type of location, ambiance, physical attributes, and people shownon the 843 video clips recorded by the participants. Five independent annotators were hired andtrained to watch the entire corpus of videos and answer 17 single and multiple choice questionson the type of location, the ambiance of the place, and characteristics of the social and physicalenvironment. The exact questions and response options are presented in Section 4.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:9
020406080100
Fig. 1. Number of videos of private places (N=301) per hour.
3.2.3 Identification of home environments. Based on the annotators’ answers to the question “Inwhat kind of place was the video clip taken?”, places were considered as “homes” in case all fiveannotators agreed on this label. In total, this procedure retained 301 videos representing homeplaces. In these environments, participants recorded videos in 64% of the cases. Reasons for notrecording a video were: “I was asked by someone not to do it” (36%), “it is not appropriate” (25%), “Idon’t feel safe” (24%) and “other” (21%). Given that participants recorded videos of the environmentwhenever they had their first drink after 8 PM or moved to another location, (i.e., change of home,or come back home, in the present case), the 301 videos illustrate home environments throughoutthe night, although with a larger proportion of those taken early in the night if the participants didnot change location. Figure 1 shows the number of videos per hour. Because of the small number ofobservations per hour after midnight, environments documented after midnight will be aggregatedin the rest of the analyses.
Due to privacy requirements requested by the Ethical Review Boards that reviewed and approvedthe project, we cannot make this dataset publicly available.
4 PHYSICAL/SOCIAL ATTRIBUTES AND AMBIANCE OF HOME SPACES (RQ1)In this section, we investigate how main patterns of physical attributes and ambiance can beextracted from videos recorded in private spaces using external annotators. In the followingsubsections, we first explain the measure (i.e. exact questions and response options), investigatethe consistency of annotations across the five annotators using Intraclass Correlation analyses, andprovide descriptive results. The Intraclass Correlation Coefficient (ICC) is a standard measure ofreliability of raters [79]. As recommended by [44], we used ICC(2,k) which is used for a fixed set ofk judges rating each target (N=301) and reflects the absolute agreement. Following the guidelinesfrom from Koo and Li [44], ICC scores below 0.5, between 0.5 and 0.75, between 0.75 and 0.9,and greater than 0.9 are indicative of poor, moderate, good, and excellent reliability, respectively.ICC(2,k) can only be computed on numerical variables, not on categorical ones, so there are fewcases in this section that does not show ICC. We summarize results of ICC scores for each possiblequestion in Table 2.
4.1 Overall representation of the spaceIn the annotation task, after carefully watching each video, several times if required, annotators wereasked to indicate “How well does the video capture the physical space (i.e. space layout, backgroundscene, furniture, decoration, etc.)?” with five single-item response options. “[1]not well at all”, “[2]notso well””, “[3]regular”, “[4]well”, and “[5]very well”. Results showed a good level of agreement onthis question (ICC = 0.83). Figure 2 shows the histogram of all individual responses of all annotatorsto this question (301 x 5 =1505). As seen in Figure 2, most of the videos were rated as providing a“regular” representation of the space. The mean of this variable is 2.91 (SD=0.86), which is sightlylower than 3 (“regular”). In some cases, participants avoided recording directly physical spaces that
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:10Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
0
100
200
300
400
500
600
700
Not well at all
Not so well
Regular Well Very well
Fig. 2. How well physical spaces are captured in videos. The y-axis represents the total number of annotations.
Physical attributes at homes ICC(2,k) mean std skewPhysical space (i.e. space layout, scene, decoration, etc) 0.83 2.91 0.86 -0.17Amount of light 0.87 2.90 0.69 -0.30How loud is the music 0.95 1.44 0.81 1.70Level of overall chatter 0.94 1.71 0.94 0.80Level of occupancy 0.97 1.82 1.09 1.12Table 2. ICC of physical attributes at homes based on N(video)=301, N(raters)=5, with scale (1-5).
could contain people. For instance, some participants recorded the ceiling or floor while panningthe camera.
4.2 Physical and Social Attributes4.2.1 Room of the home. In the annotation task, the room type within the home was labeled usingthe question: “Where in the home was the video taken?” and the following single-item response op-tions: “[a]living room”, “[b]dining room”, “[c]kitchen”, “[d]bedroom”, “[e]corridor”, “[f]terrace/balcony”,“[g]other”, and “[h]impossible to say”.
Figure 3 shows the frequency with which individual annotators identified specific rooms of thehomes in the 301 videos. Living room, bedroom, kitchen, and dining room are the most attendedspaces within homes. This result echoes previous work that using traditional methods reportedthat living rooms and bedrooms are the most used places in small and large homes by occupants[43], and extends this previous finding by showing that for the specific case of young people onweekend nights, kitchens and dining rooms are also frequently used indoor spaces. As mentionedpreviously, a few videos avoid capturing directly the physical spaces by turning the camera to theceiling and floor. This is one of reasons why “Impossible to say” appears in Figure 3.
4.2.2 Brightness. The annotators were asked to answer a single choice question “Describe theamount of light in the place” with five choices “[1]It is very dark”, “[2]It is quite dark”, “[3]Normal”,“[4]It has a good lighting” to “[5]Is is very bright”. The ICC(2,k) of brightness is high (0.87). Thebrightness variable has a mean slightly below the middle of the scale (2.9, SD=0.86). Figure 4 showsthe histogram of annotated brightness, brightness per hour (8:00-8:59 PM, etc.), and brightness perhour expressed as a percentage within that timeslot, respectively. The percentage of darkness (quitedark and very dark) increases from 18% (8PM) to 35% (0-3AM) in Figure 4c. Conceptual work ingeography [78] has recently discussed how individuals at home in the dark might be more willingto open themselves to others, and how adjusting the darkness of the home environment can be
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:11
0100200300400500600
Fig. 3. Types of spaces at homes captured in videos. The y-axis represents the total number of annotations.
a)0
100200300400500600700800900
very dark quite dark
Normal good lighting
very bright b)
0
100
200
300
400
500
20 21 22 23 0-3
very dark quite dark Normal
good lighting very bright c)
0%
20%
40%
60%
80%
100%
20 21 22 23 0-3
very dark quite dark Normal
good lighting very bright
Fig. 4. (a) Brightness, (b) Brightness per hour (8:00-8:59 PM, etc), (c) Brightness per hour expressed as apercentage within that timeslot. The order of levels of brightness (very dark, quite dark, normal, etc.) isleft-to-right in graph a, and top-to-bottom in graphs b and c. The x-axis on graph b and c is hour on Friday andSaturday nights from 20:00 to 3:00. The y-axis on graphs a and b represents the total number of annotations,while the y-axis on graph c represents the percentage normalized on each hour.
empowering. Our annotations suggest that as the weekend night goes on, young people at homeindeed tend to be in conditions of lower illumination. As a reminder, note that given the season ofthe year when the data was collected (mid September through December), it was past sunset timeat the beginning of each recorded night (8PM).
4.2.3 Music Loudness. Regarding music loudness at home places, the annotators were asked toanswer “Describe how loud is the music in the place” with five choices “[1]No music”, “[2]Low”,“[3]Medium”, “[4]Loud”, and “[5]Very loud”. The ICC(2,k) of music loudness is excellent (0.95). Themean value (1.44) is low (SD = 0.81). The skew is large (1.70) showing that the distribution has a tail.Figure 5 also shows the corresponding temporal trends. Overall, the present results on music andbrightness levels at home are consistent with recent ethnographic research showing that youngpeople tune their home by turning off lights and choosing slow paced music when they spendtime drinking with their friends at night [94]. We found that no music was played in most of therecorded environments (frequency: 76%; see Figure 5a). When music was played, the loudness levelwas quite low throughout the night (Figure 5c), suggesting that the cohort of young people arerelatively quiet in their private nightlife.
4.2.4 Chatter Loudness. The annotators were asked to describe the level of chatter loudness athome space by answering the question “Describe how loud is the chatter in the place” with five singlechoices “[1]No chatter”, “[2]Low”, “[3]Medium”, “[4]Loud”, and “[5]Very loud”. Similarly to musicloudness, the ICC agreement for chatting loudness is very high (0.94). The mean value is low (1.71,SD = 0.94). Figure 6a-c shows that there is not much loud talking in the recorded videos. Relative to
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:12Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
a)0
200
400
600
800
1000
1200
no music low medium loud very loud b)
0
100
200
300
400
500
20 21 22 23 0-3
no music low medium loud very loud c)
0%
20%
40%
60%
80%
100%
20 21 22 23 0-3
no music low medium loud very loud
Fig. 5. (a) Music loudness, (b) Music loudness per hour, (c) Music loudness per hour expressed as a percentagewithin that timeslot. The x-axis on graph b and c is hour on Friday and Saturday nights from 20:00 to 3:00. They-axis on graphs a and b represents the total number of annotations, while the y-axis on graph c representspercentage normalized on each hour.
each hourly slot, medium and loud chatting slightly increase from 8 PM to 11 PM (Figure 6c). Thisresult is clearly connected to the results obtained for the occupancy of the physical space discussednext.
4.2.5 Occupancy. Annotators were also required to describe the level of occupancy of the placeby using the following single choice question “Describe the level of occupancy of the place based onwhat you hear or see” with five choices “[1]Empty”, “[2]There are few people for this space”, “[3]It’shalf empty/half full”, “[4]It’s well attended, but there could still be more people” to “[5]It’s highlycrowded/packed”. Annotator agreement of occupancy level was excellent (ICC = 0.97). The mean oflevel occupancy of the place is 1.82 (SD=1.09). While we anticipated that most young people meetwith others at home on weekend nights, Figure 6d shows that empty is the most common category.Figure 6f also shows that young people slightly reduce gathering together from 8 PM to 10 PM;then, gathering increases at 11 PM, and decreases again after midnight.
4.2.6 Number of people present. As a complement to occupancy, we asked the annotators “Howmany people appear on the video (in addition to the phone holder)” with six choices “[1]0 (the personseems to be alone)”, “[2]1”, “[3]2-4”, “[4]5-10”, “[5]More than 10” to “[6]Impossible to say”. Figure 6gshows that around 40% of videos are labeled as containing no people, which is consistent with thelabeling of occupancy.
4.2.7 Gender of people present in videos. Among the people present in the videos, we examined theirgender ratio by asking one single choice question: “What is the gender ratio of the relatives, friends,or acquaintances appearing in the video?” with 6 response options “[1]Women only”, “[2]Mostlywomen”, “[3]Half-half”, “[4]Mostly men”, “[5]Men only” to “[6]Impossible to say”. Figure 7a showsthat “men only” is the most common situation, followed by “women only” and “half-half”. Thetotal number of situations with “men only” and “mostly men” is higher than those with “womenonly” and “mostly women”, suggesting that men appeared more often in the videos than women.Surprisingly, the 301 videos were fairly evenly distributed per gender, with 144 videos recorded by52 male participants and 157 videos recorded by 50 female participants. Figures 7b and 7c show thegender repartition of the people present in the videos recorded by male and female participants,respectively. Male participants mostly tend to spend their nights at home with other male friendsand less so with women, while no clear preference could be observed for female participants. As apoint of reference, work on a sample of 377 students [93] showed that young females tend to hangout at home with friends more than males do.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:13
a)0
100200300400500600700800900
no chatter
low medium loud very loud
b)
0
100
200
300
400
500
20 21 22 23 0-3
no chatter low medium loud very loud c)
0%
20%
40%
60%
80%
100%
20 21 22 23 0-3
no chatter low medium loud very loud
d)
0100200300400500600700800900
e)
0
100
200
300
400
500
20 21 22 23 0-3
Empty few people
half empty/full well attended
highly crowded f)
0%
20%
40%
60%
80%
100%
20 21 22 23 0-3
Empty few people
half empty/full well attended
highly crowded
g)
0100200300400500600700800
h)
0
100
200
300
400
500
20 21 22 23 0-3
0 1 2-4 5-10 >10 Impossible to say i)
0%
20%
40%
60%
80%
100%
20 21 22 23 0-3
0 1 2-4 5-10 >10 Impossible to say
Fig. 6. Annotation of (a-c) chatter level, (d-f) occupancy level, (g-i) the number of people in the videos. Theleft column shows the overall trend, the middle column the trend per hour, and the right column the relativepercentage for each timeslot. The order of values of all legends is left-to-right, top-to-bottom in all graphs.The x-axis on graphs b-c, e-f, h-i is hour on Friday and Saturday nights from 20:00 to 3:00. The y-axis ongraphs a-b, d-e, g-h represents the total number of annotations while the y-axis on graph c, f, i representspercentage normalized on each hour.
4.2.8 Activities of people. In order to assess the activities of young people at their home spaces, weasked annotators to indicate “What things are people doing in the video?” with 14 multiple choicesitems shown in Figure 8a. Results showed that activities are quite diverse, with drinking, chatting,watching TV, using smartphone/tablet/computer, and eating as the five most common activities. Asseen in Figure 8c, these main activities are roughly constant from 8 PM to midnight. Drinking, asthe most commonly annotated activity, takes 15-25 percent in relative terms across all hourly slots.The prevalence of this activity at home is not surprising given that participants were requestedto document the environment when they had their first alcoholic or non-alcoholic drink there.Nevertheless, this finding also echoes to previous research from Valentine et al. [89] showing that73% of young people report having consumed alcohol at their homes and 64% at their friends’houses over the last year. Yet, our analysis brings a finer grained description of temporal trends. Inaddition, we also examine activities of young people depending on the level of occupancy and typeof space at homes, as shown in Figure 9. In Figure 9a, when the place is empty (i.e. only the personrecording the video is present), the most commonly annotated activities are watching TV, using acomputer/tablet/smartphone and, to some extent, drinking. Conversely, in the presence of otherpeople, the commonly annotated activities are chatting, drinking, and eating, whose proportions
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:14Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
a)
0
50
100
150
200
250
b)
00.05
0.10.15
0.20.25
0.30.35
0.40.45
Women
only
Mostly
women
Half-ha
lf
Mostly
men
Men on
ly
Impo
ssible
to sa
y
c)
00.050.1
0.150.2
0.250.3
0.350.4
0.45
Women
only
Mostly
women
Half-ha
lf
Mostly
men
Men on
ly
Impo
ssible
to sa
y
Fig. 7. (a) Frequency of gender of people appearing in 301 videos recorded by 102 male and female participants,(b) percentage of gender of people appearing in 144 videos recorded by 52 male participants, and (c) percentageof gender of people appearing in 157 videos recorded by 50 female participants. The y-axis on graph a representsthe total number of annotations, while the y-axis on graphs b and c represents percentage normalized oneach possible value on the x-axis.
increase along with levels of occupancy. It might also be noticed that playing board games was themost frequently reported in “half empty/full” homes, and some dancing was reported in highlycrowed homes. Figure 9b shows that there are four places at home spaces that co-occur withspecific activities: terrace/balcony/corridor; kitchen/dining room; living room; and bedroom. Inrelated CSCW work, Baillie et al. [9] study leisure (private) and leisure (public) places in terms oftheir utility to inhabitants of a house. We complement this by showing that chatting and drinkingoccur more (in distributional terms) in leisure public areas within homes (terrace/balcony/corridor,kitchen/dining room, living room), while activities like using computer/tablet/smartphone andwatching TV occur around 60% in private leisure spaces (bedroom).
4.2.9 Reactions of people around in videos. To conclude our research on physical and socialattributes at home spaces, we examined reactions of people around in videos by asking fiveannotators to answer a single choice question “Can you see or hear one or more persons reacting toor being aware of the video being recorded?” with two answers “[1]Yes” and “[2]No”. If the previousquestion gets answered “Yes”, we will ask five questions listed in Figure 10b with three singlechoices “[1]Yes”, “[2]No”, “[3]Not sure”We are interested in how people in videos react to video recording in home spaces. As we
mentioned, many videos did not get recorded by design, as participants were told not to do it if notappropriate. Regarding the 301 recorded videos at home spaces, in 25% of cases did people in thevideo react to the camera (shown in Figure 10a). Two of the main reactions were having fun whilethe video is recorded and asking about or commenting on the purpose of the video. It is importantto note that participants in the study were explicitly instructed to record video only when it was
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:15
a)
0
100
200
300
400
500
600
b)
0
200
400
600
800
20 21 22 23 0-3
Eating Drinking
Chatting Watching TV
Using computer/tablet/smartphone Playing drinking games
Playing other games Getting ready to go out
Travelling / Walking Dancing
Resting Killing time / hanging around
Other Impossible to say
c)
0%
20%
40%
60%
80%
100%
20 21 22 23 0-3
Eating Drinking
Chatting Watching TV
Using computer/tablet/smartphone Playing drinking games
Playing other games Getting ready to go out
Travelling / Walking Dancing
Resting Killing time / hanging around
Other Impossible to say
Fig. 8. Frequency of occurrence of (a) activities (b) activities per hour, and (c) percentage of activities withineach timeslot. The order of activities (eating, drinking, chatting, watching TV, etc.) is left-to-right in graph a,and top-to-bottom in graphs b and c. The x-axis on graph b and c is hour on Friday and Saturday nights from20:00 to 3:00. The y-axis on graphs a and b is the total number of annotations while the y-axis on graph c ispercentage normalized on each hour.
socially acceptable and agreed and they were free to avoid recording [72]. The video dataset usedhere was recorded with such guidelines. There are just a few cases showing that people in the videowere not comfortable about being recorded or to hide their face.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:16Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
a)
0% 20% 40% 60% 80% 100%
Emptyfew people
half empty/fullwell attended
highly crowded
Chating EatingDrinking Watching TVUsing computer/tablet/smartphone Playing drinking gamesPlaying other games Getting ready to go outTravelling / Walking DancingResting Killing time
b)
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Terrasse / balcony
Corridor
Bedroom
Kitchen
Dining room
Living room
Chatting EatingDrinking Watching TVUsing computer/tablet/smartphone Playing drinking gamesPlaying other games (cards, video games, etc) Getting ready to go outTravelling / Walking DancingResting Killing time / hanging around
Fig. 9. Percentage of activity based on a) level of occupancy, and b) type of space at homes. The order ofactivities (chatting, eating, drinking, watching TV, etc.) is left-to-right, top-to-bottom in graphs a and b.
a)
381
1124
yes no b)
0 50 100 150 200 250 300 350 400
At least one person has fun while the video is recorded
At least one person makes fun of the person recording
At least one person hides his/her face
At least one person is not ok about being recorded
At least one person asks about or comment on the purpose of the video
Yes No Not sure
Fig. 10. (a) Yes/No reactions of people in videos. (b) Description of reacrions to the videos.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
Table 3. ICC of ambiance categories at homes based on N(videos)=301, N(raters)=5 with scale (1-7).
4.3 Ambiance attributesTo assess the ambiance of home environments , we used a modified version of the Personal LivingSpace Cue Inventory (PLSCI) [29]. This instrument was originally designed to describe personalliving spaces, e.g. rooms in family households, dormitories, or residential places. In our case, weaugmented the PLSCI with ambiance attributes from previous work [31], [74], [60], [67]. As a result,we obtained a list of 11 ambiance word groups (e.g. large/spacious, cramped/confined; all items arelisted in Table 3). A Likert scale, used in previous ambiance work but also as a reliable methodologyto annotate image aesthetics [80] was used in our work. Annotators had to rate each ambiance byindicating, on a 7-point Likert scale ranging from “[1]strongly disagree” to “[7]strongly agree”, thedegree to which they agreed with each of the ambiance attributes.
As seen in Table 3, moderate-to-good agreement levels were found for 8 out of the 11 ambiancecharacteristics (ICC greater than 0.5), but 3 items, namely festive/fun, serious/boring, and off-the-beaten-path/unique had ICC under 0.5. Attributes relating to physical characteristics of the place(large/spacious, cramped/confined) and its brightness (dark/badly-lit, bright/well-lit) have the highestagreement ranked as good (between 0.75 and 0.9). This indicates that the ambiances relating tophysical attributes are easier to rate than attributes relating to the annotators’ judgments on moresubjective variables (Serious, boring, Festive, fun and Off-the-beaten-path/unique). This result is inconcordance with the work of Nguyen et al. [60] on Airbnb personal homes, in that annotation onambiance requires observers to make abstract impressions, which makes consistent annotationchallenging for variables like festive/fun, serious/boring and Off-the- beaten-path/unique. Regardingsdescriptive statistics, the highest mean values are obtained for dull/simple (4.12), colorful/decorated(4.02), large/spacious (3.89), and dark/badly-lit (3.82).
4.3.1 Ambiance Correlation. Table 4 displays the Pearson correlation between the annotatedambiances for all home places (N=301). In Table 4, we only show correlation above 0.20 and p-value<0.001. From this analysis, we can identify opposing pairs, e.g. large/spacious vs. cramped/confined,and dark/badly-lit vs. bright/well-lit. but also observe other effects. All characteristics are associatedwith some others, with clearly identifiable patterns. First, characteristics related to brightness,namely dark/badly-lit and bright/well-lit, are uncorrelated to all other ambiance characteristics,suggesting that variations in lightings are independent of the general perceived ambiance. Second,characteristics of serious/boring, cramped/confined, and dull/simple were all grouped together (i.e.,positive correlations between all three characteristics), while characteristics of large/spacious,
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
[k] Serious, boring -Table 4. Pearson correlation of ambiance (based on N(video)=301 having p-value <0.001). Entries markedwith (*) correspond to correlation <0.20 and p-value >0.001.
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
serious, boringoff-the-beaten-path, unique
sophisticated, stylishfestive, fundull, simple
comfortable, cozybright, well-lit
cramped, confinedcolorful, decorated
dark, badly-litlarge, spacious
Drinking Chatting Watching TVUsing computer/ tablet/ smartphone Eating Playing other games (cards, video games, etc)Playing drinking games Getting ready to go out Travelling / WalkingDancing Resting Killing time / hanging around
Fig. 11. Percentage of all activities co-occur with all ambiances. The order of activities (left-right and top-bottom on the legend), e.g., drinking, chatting, watching TV, using computer/tablet/smartphone, etc. areplotted from the left (0%) to the right (100%) on the stacked bar.
colorful/decorated, comfortable/cozy, sophisticated/stylish, off-the-beaten-path/unique, and festive/funwere also grouped together.
4.3.2 Co-occurrence of Ambiance and Activities. Figure 11 shows the relative distribution of ac-tivities for the different types of ambiances. For the figure, each ambiance was binarized, suchthat each place is associated to a given ambiance only if the average rating over all annotatorsis above the mean scale (4.0). Overall, ‘Drinking’, ‘Chatting’, ‘Watching TV’, ‘Using computer/tablet/ smartphone’ and ‘Eating’ were the most prevalent activities, independently of the ambiance,although subtle variations can be observed. For example, chatting was more prevalent in unique,large, and sophisticated places, while the use of electronic devices seemed more prevalent in serious,dull, and confined places. The only ambiance that seemed largely different from the others is festive,fun, which showed a lower proportion of watching TV and using electronic devices than the otherambiances.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:19
4.4 Automatic extraction of audio and visual descriptors of home environments4.4.1 Video Preprocessing. We extract visual and audio descriptors of places from the 301 10-secondvideo clips using deep learning. Table 5 summarizes the outcomes of the learning models presentedbelow. Following the recommendation in [42] to extract at least 8 frames per second using uniformsampling, we extract a total of 29K frames. Meanwhile, we also extract 301 audio files for all videosby using command line FFmpeg [26].
4.4.2 Object Parser. To obtain an object-level description for each video, we used a deep learningmodel to extract the probability of object appearance in each frame. We applied the Inception-v3model [83] trained on the ImageNet Large Visual Recognition Challenge. This model classifiesentire images into 1000 classes (e.g. dishwasher, refrigerator, etc.) where the output for each imageat the last layer is the probability distribution over all object classes (i.e., the sum of the scores overthe 1000 classes is 1.0). The work in [83] presented the fraction of test images for which the correctclass label is not among the top five labels identified by the algorithm, namely “top-5 error rate”,reported to be 3.46%. As a result, for each frame, we have a 1000-dimensional vector with eachelement as a probability. Then, we aggregate them at the video-level over all frames to include allthe existing objects by computing, for each class, the maximum probability over the set of videoframes.
4.4.3 Scene Parser. To obtain a scene-level description for each video, we extract 365 place classes(e.g. kitchen, living room, etc) using Resnet18 [98] trained on the Place-365 database [97] for eachframe. The semantic categories of the place classes are defined by their function, e.g., dressing roomfor dressing, locker room for storing, etc. As explained online, the database is meant to be used for“high-level visual understanding tasks, such as scene context, object recognition, and action andevent prediction.” The output of the last layer is a 365-dimensional vector in which the sum of allelement values is 1. In order to represent the scene of the full video, we aggregate vectors over allframes of each video by computing the average for each class.
4.4.4 Sound Parser. To get a scene-level representation of the sounds present in a video, we extract527 audio classes using Vggish trained on the Audio Set dataset of generic audio events, which has1.7 million human-labeled 10-second YouTube video soundtracks [27]. The output of the last layeris the probability of each individual sound detected by the model.Figure 13a, b, c shows the top 30 descriptions extracted for 1000 objects, 365 places, and 527
sound classes, respectively. Overall, most of the identified top objects (e.g. TV, closet, sliding door,etc.), places (e.g. dorm, closet, etc.) and sound (e.g. speech, music, etc.) clearly correspond to homeenvironments. This said, a few unexpected results are worth commenting. First, the first placeobtained by category “jail cell” in Figure 13b seems strange. However, manual inspection of theseimages shows that studios with shelves or small rooms can indeed be mistaken with jail cells.In order to illustrate the kind of content of the Y@N video dataset, we plot four pictures in
Figure 12, with the first row as examples of good recognition, and the second row as examples ofpartly incorrect recognition. For privacy reason, the original Y@N video content cannot be shown.Second, only two sounds (music and speech) are often identified in the audio tracks, while no othersounds seem to be typical to home contexts on weekend nights.
In summary, this section answers RQ1 (consistency of annotation and the main findings from theannotation results and machine-extracted features). In each section of physical/social attributes andambiances, we present measures, ICC, and main findings . The ICC(2,k) shows that ambiance andphysical/social attributes at home (e.g., presentation of home spaces, brightness, music loudness,chatter loudness) can be consistently annotated by external observers. The results also revealthat living room, dining room, kitchen, and bedroom are common places at home where nightlife
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:20Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
a) b)
c) d)
Fig. 12. Illustration of similar content to Y@N videos. We use example photos from Pixabay [2] with PixabayLicence [1] (instead of original examples from Y@N) for privacy reasons: image a by JayMantri [41], imageb by JamesDeMers [39], image c by RonPorter [70], and image d by viganhajdari [91]. Images (a-b) areexamples of good recognition. They contain the top-5 detected CNN features from 1000-object classes (forimage a: (‘quilt, comforter, comfort, puff’, ‘studio couch, day bed’, ‘wardrobe, closet, press’, ‘sliding door’,‘four poster’); for image b: (‘table lamp’, ‘studio couch, day bed’, ‘china cabinet, china closet’, ‘lampshade,lamp shade’, ‘four poster’); and 365-place classes (for image a: (‘youth hostel’, ‘dorm room’, ‘bedroom’, ‘berth’,‘hotel room’); for image b: (‘living room’, ‘television room’, ‘waiting room’, ‘home theater’, ‘beauty salon’).Images (c-d) are examples of partly incorrect recognition. They contain the top-5 detected CNN featuresfrom 1000-object classes (for image c: (‘crate’, ‘safe’, ‘chest’, ‘cradle’, ‘carton’); for image d: (‘china cabinet,china closet’, ‘toyshop’, ‘thimble’, ‘bookcase’, ‘medicine chest, medicine cabinet’, ‘tobacco shop, tobacconistshop, tobacconist’); and 365-place classes (for image c: (‘jail cell’, ‘burial chamber’, ‘dorm room’, ‘bedchamber’,‘stable’); for image d: (‘bookstore’, ‘gas station’, ‘toyshop’,‘library/indoor’, ‘storage room’). The top-5 detectedobjects and scenes of images a-b are more relevant, while the recognized content of images c-d is moreirrelevant.
activities like eating/drinking, entertainment (watching TV or using mobile devices) and chattinghappen. Young people at home weekend nights seem to be mindful about the loudness of musicand level of chatter. In addition, we found a surprisingly large proportion of videos with no peopleother than the volunteer, engaged in relatively quiet activities. Although the number of videoscontains people do not take a large portion, they describe the gender ratio and their activities aswell as their reactions to our participants. Moreover, although there are still unexpected resultsof extracted objects and scenes, many identified CNN-extracted classes from objects, scenes, andsounds are relevant to home environments. To our knowledge, this analysis of nightlife activitiesat home, which was enabled by the crowdsensing experience, has not been previously reported.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:21
FeatureClasses
Frame Level (28Kframes) Video Level (301 videos)
1000classes
Probability distribu-tion over 1000 objectclasses (Sum of 1000classes is 1)
Class-specific aggregate for each video: maximum prob-ability over the set of frames for each class.Purpose: obtain a representation of the objects presentin the video.
365 scenecategories
Probability distribu-tion over the 365scene classes (Sum of365 classes is 1)
Class-specific aggregate for each video: average proba-bility over the set of frames for each class.Purpose: obtain a representation of the most likely scenein the video.
527sounds Not available Probability distribution over the 527 sound classes
Table 5. Visual and sound extracted features for the video dataset.
5 MACHINE-EXTRACTED FEATURES AND AMBIANCE RECOGNITION OF HOMESPACES (RQ2)
This section describes how machine analysis of the audio-visual tracks of videos can be used tocharacterize and enrich the understanding of youth home spaces on weekend nights.
5.1 Correlation between Machine-extracted Features and AmbianceThis section aims to identify what machine-learning extracted features (1000-object classes, 365-place classes, and 527-sound classes) are correlated with the 11 ambiances categories assessed bythe annotators.
5.1.1 Correlation between ambiances and object classes. Correlation results with ambiance areshown in Table 6. Only the largest correlations are shown, (i.e. those higher or equal to 0.25 andwith p-value < 0.001). Places described as comfortable/cozy have couches and beds present in thevideos, while festive/fun places were positively correlated with eating places and movie places.These results were confirmed by manual inspection of the videos. We also noted that, in a fewcases, participants recorded the TV program they were watching as part of their home space videos.This might explain why dark ambiances are correlated with objects like cinema, but also withseemingly random objects like car mirror or grey fox. This is a known limitation of using CNNmodels trained on datasets which are not specifically designed for home environments [83]. Thiscould make some unexpected objects recognized and associated. Interestingly, object category“restaurant, eating house, eating place, eatery” has a positive association with festive/fun ambiance,while has a negative correlation with dull/simple, and serious/boring ambiance.
5.1.2 Correlation between ambiances and scene classes. Correlations between the 365-scene classesand the 11 ambiances categories are shown in Table 7. Overall, the results show similar associationsto those identified in Table 6. For example, a bedroom and a living room associate positivelywith comfortable ambiance. A dining hall and dining room are positively linked to large/spaciousambiance, while pantry or closet do with cramped/confined and dull/simple ambiance. Results alsoshow that dark and bright ambiances are correlated, negatively and positively, with a large numberof scene classes. As mentioned above, participants have sometimes recorded videos of TV programsin dark places at homes, which made the model recognize some places types erroneously, i.e., theplaces depicted on the TV shows; recall that watching TV was a very popular activity (Figure 8).
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:22Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
a) 0 0.05 0.1 0.15 0.2 0.25
television, television systemwardrobe, closet, press
sliding doorhome theater, home theatre
medicine chest, medicine cabinetrefrigerator, icebox
Fig. 13. Top 30 features of (a) 1000 object classes, (b) 365 place classes, and (c) 527 sound classes.
5.1.3 Correlation between ambiances and sound classes. Finally, we examine the correlation betweenthe 527 sound features with ambiances. Only three correlations were above 0.2 with p-value <0.001.In particular, festive/fun ambiance are positively associated with chewing/mastication (r = 0.21),which might be explained as the same ambiance categories were associated to people eating (seeTables 6 and 7). Also, the correlation of female singing (0.20) and techno (0.20) (both music-relatedsounds) with off-the-beaten-path/unique ambiance could help explain why that ambiance has
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:23
Ambiance Features 1000-object classes[a] Large, spacious medicine chest, medicine cabinet(-0.33), refrigerator, icebox(-0.31)
[b] Dark, badly-lit
cinema, movie theater, movie theatre, movie house, picturepalace(0.37) , grey fox, gray fox, Urocyon cinereoargenteus(0.32), suspension bridge(0.30) , hyena, hyaena(0.28) , wing(0.25), bad-ger(0.27) , miniature pinscher(0.26), jack-o-lantern(0.26), desktop com-puter(0.26), car mirror(0.25)whiptail, whiptail lizard(-0.28), microwave, microwave oven(-0.28)
[h]Festive, fun restaurant, eating house, eating place, eatery(0.26), cinema, movietheater, movie theatre, movie house, picture palace(0.25)
[i]Sophisticated, stylish *[j]Off-the-beaten-path,unique lumbermill, sawmill(0.31), dam, dike, dyke(0.27)
[k]Serious, boring restaurant, eating house, eating place, eatery(-0.26)Table 6. Pearson correlation between ambiance and 1000-object classes limited to classes with Pearsoncorrelation score >=0.25 and p-value <0.001. Negative and positive correlation values are ranked in descendingorder by absolute correlation value and are shown in red and blue, respectively. Entries marked with (*)correspond to p-value >0.001 and are not discussed.
positive correlations in home environments with features like bedchamber (0.42), throne room(0.37), or living room (0.33) in Table 7.
5.2 Ambiance InferenceThis section presents the investigation of whether and how the ambiance of home places can beautomatically inferred using machine-extracted features.
5.2.1 Inference task, method, and performance evaluation. The goal is to infer (in the regressionsense) the ambiance of home spaces as perceived by external observers. This inference task uses theaggregated annotations of ambiance discussed in previous sections and is run on the video, whichis aggregated as described in the previous section. Random Forest (RF) [15] is used as a regressionmodel in our inference task. By using RF, multiple decision trees are built up to form variousclassification outputs. In this experiment, we set parameters ntrees = 500 as recommended by [50].We ensure that the train and test set take 80% and 20%, respectively. We also apply 5-fold crossvalidation for training phase. After obtaining RF trained models, we quantify the performance by
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
Table 7. Pearson correlation between ambiance and 365-scene classes limited to classes with Pearson correla-tion score >=0.25 and p-value <0.001. Negative and positive correlation values are ranked in descending orderby absolute correlation value and are shown in red and blue, respectively. Entry marked with (*) correspondsto p-value >0.001 and Pearson correlation score <0.25.
Table 8. Pearson correlation between ambiances and 527-sound classes with Pearson score >=0.20 and p-value<0.001. Entries marked with (*) correspond to p-value >0.001 and are not discussed.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
Table 9. Inference results including Pearson’s correlation coefficient (r), coefficient of determination (R2). AllR2 with score >= 0.20 are shown in bold font. Rows marked with (*) correspond to ambiance categories thatdid not reach sufficient annotator agreement (ICC).
using Pearson’s correlation coefficient (r), and the coefficient of determination (R2). In the contextof our RF model, R2 measures how much variance in ambiance is explained by the RF model.
5.2.2 Experiment and results. We randomly divide the 301 videos into two subsets: 80% (241 videos)for training and 20% (60 videos) for testing. We apply RF on 241 videos for training with 5-fold crossvalidation. The evaluation of RF model is shown in Table 9. We observe that the audio features arenot capable of improving over a simple prediction of the mean score (R2 ∼ 0). In contrast, using1000 object classes can infer certain ambiances of home spaces with R2 > 0.2, namely large/spacious,dark/badly-lit, bright/well-lit, and dull/simple. The highest R2 obtained is 0.44 for bright/well-lit.Meanwhile, the rest of ambiance categories cannot be inferred by the object representation. Recallthat three of these ambiance categories (festive/fun, serious/boring, off-the-beaten-path/unique) hadnot reach sufficient ICC agreement (Table 3), but we decided to include the results for purposes ofcompleteness. Regarding the 365-scene classes, five of the eleven ambiance variables (large/spacious,dark/badly-lit, bright/well-lit, cramped/confined, and comfortable/cozy) are predicted by using 365-scene classes with R2 > 0.2 (R2 = 0.69 for dark/badly-lit). In Section 5, we discussed the correlationof ambiances and scenes. Clearly, certain scenes can predict those ambiances related to spacecapacity (large/spacious vs cramped/confined), and brightness (bright/well-lit vs dark/badly-lit). Forcomfortable/cozy ambiance, Section 5 also showed that living room with couch, and bedroom withbed, have positive correlation. Interestingly, two of the ambiance variables (colorful/decorated andsophisticated/stylish) could not be inferred by any of the visual representations, regardless of thefact that they achieved good inter-annotator reliability, (0.66 and 0.65, respectively, see Table 3).
In summary, we use RF to train a regression model and use R2 as the main measure to evaluatewhich features can predict the ambiance of a home space. Our findings show that six of the ambiancecategories can be inferred with R2 in the [0.21, 0.69] range (four with object-based features, andfive with scene-level features), and with higher R2 values when a scene deep network is used.More specifically, space capacity (large/spacious vs cramped/confined), brightness (bright/well-lit vsdark/badly-lit), comfortable/cozy, and dull/simple can be predicted by object-level and scene-leveldescription. In contrast, audio features were not effective at inferring ambiance.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:26Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
6 DISCUSSION AND IMPLICATIONSTable 10 summarizes our main findings for RQ1 and RQ2. We now discuss the results and some oftheir implications.
6.1 A Unique video Dataset of Home SpacesIn terms of data source, collecting data on home environments via crowdsourced videos is novel incomparison to previous work using social media sources. This includes research on Foursquare,which showed that users underreported home presence by checking into homes considerably lessfrequently than into other places, given the logic of such social network [55], [22]; and also includesrecent work on Airbnb, which is known to feature photos of homes that are taken with the explicitpurpose of attracting possible guests, in some cases taken by professional photographers [60]. Ourstudy used 301 ten-second video clips of young people’s home spaces on weekend nights. To ourknowledge, this is a unique dataset of real-life home environments that cannot be compared to anyother publicly available dataset, in the sense that participants’ showed their home spaces simply asthey are (with no artistic filters or advertising intentions) on their weekend nights.
From the total set of 843 videos collected in the study, slightly less than one third were consistentlyidentified as representing homes by the five annotators. Yet, this does not mean that only one-thirdof the nights were spent at home, but rather it can be seen as a consequence of the study design,which requested participants to provide only one video per night if they did not change locationduring the night. Given that about half of all drinks (non-alcoholic or alcoholic) in the Youth@Nightdataset were documented in homes [48], this result suggests that participants were less likely tochange locations when starting the night at home than when going out [46], highlighting therelevance to research and understand what happens in this usually hidden or hard-to-reach kind ofenvironment. In addition, the levels of inter-annotator agreement for most of the physical attributesat homes were globally good to excellent. This result echoes previous work in psychology [30] thatfound that personal environments elicit similar impressions from independent observers, whileadding the novel angle of using short video as stimuli (rather than photos). This result also indicatesthat, despite being relatively short, 10-second videos are long enough to provide adequate cues ofthe physical and social environment, the ongoing activities, and the ambiance.
6.2 Home as a Nightlife SpaceAs mentioned above, about one third of the Y@N videos were recorded in homes, and participantswere less likely to change locations when starting the night at home than when going out. Thishighlights the need to understand this particular environment. Qualitative feedback from theparticipants at the end of the fieldwork echoed previous research that has found that homes canserve both as ‘prequel night out spaces’, where young people meet, dress up, and get ready forthe night out, as well as a standalone nightlife space where they hang out with friends or haveparties [51]. For one participant, home was his main nightlife destination: “Now that I study inLausanne and live here, when I go out it’s really to other people’s place or at my place. Which stilldoes not prevent me from going out [to pubs and clubs] now and then”. Another participant mainlyconceived home as the starting point of the night: “Well, when I go out, I prefer drinking beforegoing out, well, not before going out but, let’s say we meet with friends and we go to someone’splace to drink or just eat and we drink something, or in a park during the summer, yeah, let’ssay I start drinking [in a residential neighborhood] and then we move on and continue the partydowntown”. Finally, several participants considered the home as an alternative to commercialnightlife venues: “For me, there are two types of nights out: the dancing ones, when we go to clubsand the point is to dance [. . . ] and then there are the quiet ones, when we just sit, at someone’s
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:27
place or in a bar, and we talk and that’s it” or “There are different kinds of nights out. Sometimes,people want to go out to meet others and that’s it, it all depends on the mood we are in that night.It’s true that sometimes we enjoy staying with friends and have big parties in homes, or go out inthe city, but as a small group.”In order to better represent home environments, the annotation task developed for this study
revealed detailed attributes of physical and social environment, including the types of roomsattended, levels of brightness, loudness and occupancy, the number and gender of people, and theongoing activities. Altogether, this information provides a comprehensive picture of young people’snightlife environments. Specifically, we examined co-occurrence between activities and levels ofoccupancy and types of spaces at home. The authors of [7], [88], [9] studied usage of domesticspaces that were used in daily life activities, and specific psychological states (e.g., mental stress).In our research, home spaces were analyzed from the perspective of activities of young peopleon weekend nights. Through physical and social attributes, we have insights of activities in thecontext of Swiss young people (16-25 year-old), who present differences to other populations, e.g.in the US, where legal drinking age and norms about the use of the public space differ from thosein Europe. We found that young people spent weekend night time watching TV, listening to music[77], and playing games. Previous findings about pre-drinking before going out in [95] or drinkingat friends’ or family’s homes [35] were also partly shown in our work.One particular instrument of the annotation task, the ambiance scale, aimed to capture the
different dimensions of this construct. Dimensions related to the physical space (e.g. large, spacious),which could be rated rather objectively by the annotators, showed a high degree of agreementamong them. Dimensions relating to the personal evaluation of the annotators (e.g. off-the-beaten-path) were indeed more subjective and showed a lower degree of agreement among the annotators.From the correlation analysis, three main groups of ambiances were identified: positively perceivedcharacteristics (large, colorful, festive, stylish, and unique), negatively perceived characteristics(cramped, simple, and boring) and independent characteristics (dark and bright). In addition, themain types of ongoing activities were consistent across ambiance categories (drinking, chatting,watching TV, and computer device use), and small variations were found, e.g. less TV watching forthe fun/festive ambiance.
While the aim of the annotation task was to describe home spaces from the perspective of humanannotators, the aim of the machine learning task was to observe home spaces through automatic-extracted features using CNNs models, [83], [98], [27]. Thus, without using external annotationof physical and social attributes, the latter task was able to automatically describe home spacesby observing the probability distribution of visual and audio labels. Correlation results betweenautomatically extracted features based on the image frames of the videos and ambiance labelsprovided promising results for the visual cues (i.e. objects or scenes) from the videos. Yet, resultsalso showed that the existing classes are made to recognize all kinds of objects or situations, evensome that are not supposed to be in homes, such as jail cell, car parts, etc. Future research is clearlyneeded for the development of a specialized dictionary of classes focused on home environments.Regarding automatic-extracted features based on the soundtrack of the videos, however, only
two of the sounds dominated the dataset (speech and music), and thus only a few associations werefound with ambiance features. These might be related to the way audio was recorded, but alsobecause homes at night are generally quiet or because not enough information was found in thesound measure in [27].
6.3 Feasibility of Ambiance InferenceWe examined the use of machine-extracted features, i.e., 527-sound, 100-object, 365-scene features,for automatic inference of ambiance. As a result, large, dark, bright, confined, comfortable, and
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:28Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
simple ambiances could be inferred by using object and scene classes. These ambiances could bealso be perceived by people while unique, festive, and boring ambiances could be ambiguous whenbeing annotated by humans. Comparing these inference results to those reported in [60] on Airbnbhome photos, our results corroborate that ambiances closer to physical attributes reach betterrecognition performance, although the performance we obtained is lower than that obtained onAirbnb data for three variables (large, comfortable, and simple), similar for one variable (confined);and higher for two variables (bright, dark). Note that in addition to the datasets being different, thespecific CNN models and the CNN outputs used as features are different too (last convolutionallayer in [60] vs. final output equal to the number of objects or scenes in our work). Note also thatwe made this choice in order to interpret the CNN-derived features in the correlation analysis inSection 5.1. For future work, we believe that regression performance could be improved by CNNadaptation, i.e., by fine-tuning the last CNN layers to the ambiance target class as demonstratedin other visual tasks [56]. Home ambiance recognizers built around short duration mobile videoscould be advantageous as they might in general contain more information than still images, andused in future applications, as discussed in the next subsection.
6.4 Implications for CSCW researchWe conclude this section by discussing some of the implications of our work for CSCW and socialcomputing research.
Understanding youth practices at home from mobile crowdsourced data. Using crowd-sourced personal videos as input, we showed that a mixed methodology combining manual annota-tion and automatically extracted features enabled an in-depth study of youth personal spaces onweekend nights with respect to physical attributes, activities, and social attributes, including jointpatterns of activities and places. Crowdsourced visual datasets like the one used here complementanother common source of data used in CSCW research, namely social media like Instagram. Whileearly research showed that the home environment was infrequently reported or talked about byusers [55], [22], future research could investigate whether certain sub-communities specificallydepict nightlife in private spaces, and what specific practices are promoted or enacted aroundthis theme, including ephemerality, self-representation, and sociality. This investigation wouldrequire the use of mixed methods of inquiry, combining machine analyses with user interviewsand surveys. Furthermore, given that the concept of nightlife is broad and encompasses both theprivate and public spheres, a second promising line of future work could investigate the interplaybetween private and public spaces in urban nightlife, and how this is expressed digitally bothin crowdsourced campaigns and social media. For instance, recent qualitative work showed thatseveral participants in the Youth@Night campaign coordinated nightlife activities via Whatsapp[85]. This research could benefit from previous CSCW literature on coordination of action andsocial participation.
Applications of home ambiance recognition.Our work on recognition of ambiance at homealso has potential implications for future CSCW work. First, it is evident from our study that state-of-art deep visual learning systems, while useful, still generate erroneous visual descriptors. Webelieve that it is important to make these limitations explicit to inform other CSCW researcherswho plan to use deep learning as a toolbox for their future work. At the same time, in a fast-movingdomain, it is not unreasonable to expect progress that could mitigate some of the current limitations,and thus to anticipate that the shown recognition performance will be improved (e.g., Facebook haspublished results on deep learning models trained on 1 billion Instagram images) [57]. With this,one could envision applications in home supporting systems. Homes are reconfigurable spaces,in which certain elements can be readily changed (decoration, spatial organization of furniture,light, and music). A system able to recognize ambiance could also make recommendations of
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
My Own Private Nightlife 189:29
suitable ambiances at home for specific activities, e.g. to promote socialization. This kind of workwould require human-centered approaches to design such prototypes, integrating perspectives ofprivacy, ethics, and transparency, all of whom are active topics of investigation in CSCW and socialcomputing [99], [8], [96], [49].
Future work on human factor in home research. In CSCW research, human factors playimportant roles requiring interdisciplinary researchers (social psychologists, sociologists, anthro-pologists, computer scientists) to find appropriate methods for an individual or group to adopttechnology into their daily activities. In this case, technology plays a supporting role while thehuman factor plays a central one. In our work, we focused on home environments and inferringhome ambiance from videos which contributes to CSCW applications, while adding to themesthat are relevant to CSCW. Besides physical and social attributes, emotional states and nightlifebehaviors and their links to ambiances could need the expertise of other researchers [63], [76].In future work, youth practices at home and ambiances, technologists could collaborate withspecialists in interior decoration art, or psychologists, to build systems to support people to linktheir home ambiances to their current emotions as well as their behaviors. Beyond building thistechnology, users would increase their self-awareness about their home ambiances and their ownbehaviors to promote positive changes and share them with others.
7 CONCLUSIONIn this paper, we presented an original study of the characteristics of night personal spaces,including manual coding of places, machine extraction of acoustic and visual description of places,and inference of ambiance of homes of young people in the weekend night setting. We conclude byrevisiting the research questions posed at the beginning of the paper.
RQ1: Given crowdsourced videos recorded at home spaces by young people at night, what patternsof physical and ambiance attributes of youth home spaces can be revealed by manual coding of videosusing external annotators and machine-extracted features? By describing measures, discussing ICC,and showing results, we sequentially analyzed the problem from physical/social attributes (homespaces, brightness, loudness, human presence, activities) to ambiances. We observed co-occurrencebetween activities and spaces at homes as well as ambiances. Then, we showed that ambiancescould be grouped into two clusters: “unlike” characteristics with serious/boring, cramped/confined,dull/simple, and “like” characteristics with Large/spacious, colorful/decorated, comfortable/cozy,sophisticated/stylish, off-the-beaten-path/unique, festive/fun. Finally, we used state-of-the-art pre-trained deep learning models to extract automatic features to represent videos, namely objects,scenes, and sounds. Most machine-extracted classes relevantly characterize home environments,but there were some unexpected features.
RQ2: What do machine-extracted features of videos reveal about physical attributes of youth homespaces? Can these machine-extracted features infer the perceived ambiance of such spaces? Correlationsbetween ambiance and automatic features potentially show the feasibility of using machine-extracted features to automatically describe home spaces, although there are certain limitations.Regarding the inference task, ambiances like space capacity (large/spacious vs. cramped/confined),brightness (bright/well-lit vs. dark/badly-lit), comfortable/cozy, and dull/simple can be inferred forprivate spaces in the weekend nights by using 1000 object classes and 365 scene classes. The totalnumber of videos (N=301) could be a limitation for mo0del training in the automatic inferenceexperiments. However, our results show that six of the ambiance categories can be inferred withR2 in the [0.21, 0.69] range, and with higher R2 values when a scene deep network is used.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
189:30Thanh-Trung Phan, Florian Labhart,
Daniel Gatica-Perez
RQ Factors Message
RQ1 -PhysicalandSocialAttributes
Home spacesThe most attended type of room is the living room; followed by bed-room; kitchen/dining room were also frequently attended rooms atnight
Brightness It tends to reduce from early night to late nightMusic loudness Videos contained no music on 76% of all situationsChatter loudness Home are mostly quiet with slight increase from 8PM to 11PMOccupancy andnumber of peoplepresent
Around 60% of videos contained people gathering from 8 PM to 11PM; then reducing after 11 PM
GenderA gender-matching pattern is evident: female participants tend togather more with other women, and conversely for male participants.Mixed groups, however, also occur.
ActivitiesDrinking, chatting, watching TV, using smartphones/computer, andeating are the most popular activities of young people on weekendnights.
RQ1 -Ambiances
Agreement on am-biance 8 of the 11 ambiance variables have ICCs above 0.5.
Correlation be-tween ambiances
Place ambiances are grouped on two main opposite dimensions,namely places seen as large, colorful, comfortable, festive, stylish,unique; versus places seen as confined, simple, boring. Dark andbright ambiances do not have a significant correlation with the restof ambiances.
RQ1 -Machine-extractedFeatures
Automatic de-scription
1000-object, 365-scene, 527-sound auto-extracted features can expressambiances but with a certain level of noise, because labels of theseclasses for CNN models are not specifically designed for homes.
Although there are some limitations on the labels of CNNs model,automatic-extracted features have reasonable correlation with am-biances.
Regression Perfor-mances
Six of the ambiance variables (large, dark, bright, confined, comfort-able, simple) can be inferred by using object and scene features withcoefficient of determination above 0.2. For the other five variables(including three with low ICC). regression performance is low.
Table 10. Summary of findings related to our two RQs.
ACKNOWLEDGMENTSThis work has been funded by the Swiss National Science Foundation through the Dusk2DawnSinergia project, and a Swiss Government Excellence Scholarship.
REFERENCES[1] [n.d.]. Pixabay Licence. Retrieved Aug 19, 2019 from https://pixabay.com/service/license/[2] [n.d.]. Pixabay Website. Retrieved Aug 19, 2019 from https://pixabay.com/[3] 2019. Definition of home in English. Retrieved June 22, 2019 from https://dictionary.cambridge.org/dictionary/english/
home#cald4-1-1-2[4] Joan Abbott-Chapman and Margaret Robertson. 1999. Home as a private space: some adolescent constructs. Journal of
youth studies 2, 1 (1999), 23–43.[5] Joan Abbott-Chapman and Margaret Robertson. 2001. Youth, leisure and home: Space, place and identity. Loisir et
société/Society and Leisure 24, 2 (2001), 485–506.[6] Akile Ahmet. 2013. Home sites: the location (s) of ‘home’for young men. Urban Studies 50, 3 (2013), 621–634.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
[7] Saeid Alitajer and Ghazaleh Molavi Nojoumi. 2016. Privacy at home: Analysis of behavioral patterns in the spatialconfiguration of traditional and modern houses in the city of Hamedan based on the notion of space syntax. Frontiersof Architectural Research 5, 3 (2016), 341–352.
[8] Karla Badillo-Urquiola, Yaxing Yao, Oshrat Ayalon, Bart Knijnenurg, Xinru Page, Eran Toch, Yang Wang, and Pamela JWisniewski. 2018. Privacy in Context: Critically Engaging with Theory to Guide Privacy Research and Design. InCompanion of the 2018 ACM Conference on Computer Supported Cooperative Work and Social Computing. ACM, 425–431.
[9] Lynne Baillie and David Benyon. 2008. Place and technology in the home. Computer Supported Cooperative Work(CSCW) 17, 2-3 (2008), 227–256.
[10] Rick Bell, Herbert L Meiselman, Barry J Pierson, and William G Reeve. 1994. Effects of adding an Italian theme to arestaurant on the perceived ethnicity, acceptability, and selection of foods. Appetite 22, 1 (1994), 11–24.
[11] Mark A Bellis and Karen Hughes. 2011. Getting drunk safely? Night-life policy in the UK and its public healthconsequences. Drug and alcohol review 30, 5 (2011), 536–545.
[12] Yassir Benkhedda, Darshan Santani, and Daniel Gatica-Perez. 2017. Venues in social media: Examining ambianceperception through scene semantics. In Proceedings of the 2017 ACM on Multimedia Conference. ACM, 1416–1424.
[13] Alison Blunt. 2005. Cultural geography: cultural geographies of home. Progress in human geography 29, 4 (2005),505–515.
[14] Danah Boyd. 2014. It’s complicated: The social lives of networked teens. Yale University Press.[15] Leo Breiman. 2001. Random forests. Machine learning 45, 1 (2001), 5–32.[16] Gulcan Can, Yassir Benkhedda, and Daniel Gatica-Perez. 2018. Ambiance in Social Media Venues: Visual Cue
Interpretation byMachines and Crowds. In Proceedings of the IEEE Conference on Computer Vision and Pattern RecognitionWorkshops. 2363–2372.
[17] Dana R Carney, John T Jost, Samuel D Gosling, and Jeff Potter. 2008. The secret lives of liberals and conservatives:Personality profiles, interaction styles, and the things they leave behind. Political Psychology 29, 6 (2008), 807–840.
[18] Paul Chatterton and Robert Hollands. 2002. Theorising urban playscapes: producing, regulating and consumingyouthful nightlife city spaces. Urban studies 39, 1 (2002), 95–116.
[19] Yohan Chon, Nicholas D Lane, Yunjong Kim, Feng Zhao, and Hojung Cha. 2013. Understanding the coverage andscalability of place-centric crowdsensing. In Proceedings of the 2013 ACM international joint conference on Pervasive andubiquitous computing. ACM, 3–12.
[20] Yohan Chon, Nicholas D Lane, Fan Li, Hojung Cha, and Feng Zhao. 2012. Automatically characterizing places withopportunistic crowdsensing using smartphones. In Proceedings of the 2012 ACM Conference on Ubiquitous Computing.ACM, 481–490.
[21] Cary C Countryman and SooCheong Jang. 2006. The effects of atmospheric elements on customer impression: thecase of hotel lobbies. International Journal of Contemporary Hospitality Management 18, 7 (2006), 534–545.
[22] Henriette Cramer, Mattias Rost, and Lars Erik Holmquist. 2011. Performing a check-in: emerging practices, normsand’conflicts’ in location-sharing using foursquare. In Proceedings of the 13th international conference on humancomputer interaction with mobile devices and services. ACM, 57–66.
[23] Jakob Demant and Sara Landolt. 2014. Youth drinking in public places: The production of drinking spaces in andoutside nightlife areas. Urban Studies 51, 1 (2014), 170–184.
[24] Cambridge Dictionary. 2018. Definition of ambiance in English. Retrieved November 11, 2018 from https://dictionary.cambridge.org/dictionary/english/ambience
[25] Cameron Duff. 2012. Accounting for context: Exploring the role of objects and spaces in the consumption of alcoholand other drugs. Social & Cultural Geography 13, 2 (2012), 145–159.
[26] Ffmpeg. 2010. FFMPEG. http://www.ffmpeg.org[27] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal,
and Marvin Ritter. 2017. Audio set: An ontology and human-labeled dataset for audio events. In Acoustics, Speech andSignal Processing (ICASSP), 2017 IEEE International Conference on. IEEE, 776–780.
[28] SD Gosling, Robert Gifford, and Lindsay J McCunn. 2013. The selection, creation, and perception of interior spaces:An environmental psychology approach. (2013).
[29] Samuel D Gosling, Kenneth H Craik, Nicholas R Martin, and Michelle R Pryor. 2005. The personal living space cueinventory: An analysis and evaluation. Environment and Behavior 37, 5 (2005), 683–705.
[30] Samuel D Gosling, Sei Jin Ko, Thomas Mannarelli, and Margaret E Morris. 2002. A room with a cue: Personalityjudgments based on offices and bedrooms. Journal of personality and social psychology 82, 3 (2002), 379.
[31] Lindsay T Graham and Samuel D Gosling. 2011. Can the ambiance of a place be determined by the user profiles of thepeople who visit it?. In ICWSM.
[32] Lindsay T Graham, Samuel D Gosling, and Christopher K Travis. 2015. The psychology of home environments: A callfor research on residential space. Perspectives on Psychological Science 10, 3 (2015), 346–356.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
[33] Tali Hatuka and Eran Toch. 2017. Being visible in public space: The normalisation of asymmetrical visibility. UrbanStudies 54, 4 (2017), 984–998.
[34] Dolores Hayden. 2004. Building suburbia: Green fields and urban growth, 1820-2000. Vintage.[35] Sarah L Holloway, Mark Jayne, and Gill Valentine. 2008. ‘Sainsbury’s is my local’: English alcohol policy, domestic
drinking practices and the meaning of home. Transactions of the Institute of British Geographers 33, 4 (2008), 532–547.[36] Mark Holton. 2016. Living together in student accommodation: performances, boundaries and homemaking. Area 48,
1 (2016), 57–63.[37] Phil Hubbard. 2008. Regulating the social impacts of studentification: a Loughborough case study. Environment and
Planning A 40, 2 (2008), 323–341.[38] Tapio Ikkala and Airi Lampinen. 2015. Monetizing network hospitality: Hospitality and sociability in the context of
Airbnb. In Proceedings of the 18th ACM conference on computer supported cooperative work & social computing. ACM,1033–1044.
[39] JamesDeMers. [n.d.]. An example of living room on Pixabay. Retrieved Aug 19, 2019 from https://pixabay.com/photos/farmhouse-bedroom-old-515619/
[40] Margaretha Järvinen and Jeanette Østergaard. 2009. Governing adolescent drinking. Youth & society 40, 3 (2009),377–402.
[41] JayMantri. [n.d.]. An example of bedroom on Pixabay. Retrieved Aug 19, 2019 from https://pixabay.com/photos/bedroom-sleeping-bed-furniture-405920/
[42] Hina Keval and M Angela Sasse. 2008. to catch a thief–you need at least 8 frames per second: the impact of frame rateson user performance in a CCTV detection task. In Proceedings of the 16th ACM international conference on Multimedia.ACM, 941–944.
[43] Iman Khajehzadeh and Brenda Vale. 2015. How do people use large houses. In Living and Learning: Research for aBetter Built Environment: Proceeding of the 49th International Conference of the Architectural Science Association 2015.153–162.
[44] Terry K Koo and Mae Y Li. 2016. A guideline of selecting and reporting intraclass correlation coefficients for reliabilityresearch. Journal of chiropractic medicine 15, 2 (2016), 155–163.
[45] Philip Kotler. 1973. Atmospherics as a marketing tool. Journal of retailing 49, 4 (1973), 48–64.[46] Florian Labhart, Kathryn Graham, Samantha Wells, and Emmanuel Kuntsche. 2013. Drinking before going to licensed
premises: An event-level analysis of predrinking, alcohol consumption, and adverse outcomes. Alcoholism: Clinicaland Experimental Research 37, 2 (2013), 284–291.
[47] Florian Labhart, Darshan Santani, Jasmine Truong, Flavio Tarsetti, Olivier Bornet, Sara Landolt, Daniel Gatica-Perez,and Emmanuel Kuntsche. 2017. Development of the Geographical Proportional-to-size Street-Intercept Sampling(GPSIS) method for recruiting urban nightlife-goers in an entire city. International Journal of Social ResearchMethodology20, 6 (2017), 721–736.
[48] Florian Labhart, Flavio Tarsetti, Olivier Bornet, Darshan Santani, Jasmine Truong, Sara Landolt, Daniel Gatica-Perez,and Emmanuel Kuntsche. 2019. Capturing drinking and nightlife behaviours and their social and physical contextwith a smartphone application - investigation of users’ experience and reactivity. Addiction Research and Theory (2019).http://infoscience.epfl.ch/record/263840
[49] Josephine Lau, Benjamin Zimmerman, and Florian Schaub. 2018. Alexa, are you listening?: Privacy perceptions,concerns and privacy-seeking behaviors with smart speakers. Proceedings of the ACM on Human-Computer Interaction2, CSCW (2018), 102.
[50] Andy Liaw, Matthew Wiener, et al. 2002. Classification and regression by randomForest. R news 2, 3 (2002), 18–22.[51] Siân Lincoln. 2012. Youth culture and private space. Springer.[52] Siân Lincoln. 2014. “I’ve Stamped My Personality All Over It” The Meaning of Objects in Teenage Bedroom Space.
Space and Culture 17, 3 (2014), 266–279.[53] Siân Lincoln. 2015. ‘My Bedroom Is Me’: Young People, Private Space, Consumption and the Family Home. In Intimacies,
Critical Consumption and Diverse Economies. Springer, 87–106.[54] Siân Lincoln and Brady Robards. 2016. Being strategic and taking control: Bedrooms, social network sites and the
narratives of growing up. new media & society 18, 6 (2016), 927–943.[55] Janne Lindqvist, Justin Cranshaw, Jason Wiese, Jason Hong, and John Zimmerman. 2011. I’m the mayor of my house:
examining why people use foursquare-a social-driven location sharing application. In Proceedings of the SIGCHIconference on human factors in computing systems. ACM, 2409–2418.
[56] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael I Jordan. 2015. Learning transferable features with deepadaptation networks. arXiv preprint arXiv:1502.02791 (2015).
[57] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, andLaurens van der Maaten. 2018. Exploring the limits of weakly supervised pretraining. In Proceedings of the EuropeanConference on Computer Vision (ECCV). 181–196.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
[58] Denise Marquard. 2014. Zurich, the Party Town. http://bit.ly/28M8nbk Accessed: July 15, 2016.[59] Hugh Matthews, Melanie Limb, and Barry Percy-Smith. 1998. Changing worlds: the microgeographies of young
teenagers. Tijdschrift voor economische en sociale geografie 89, 2 (1998), 193–202.[60] Laurent Son Nguyen, Salvador Ruiz-Correa, Marianne Schmid Mast, and Daniel Gatica-Perez. 2018. Check out this
place: Inferring ambiance from airbnb photos. IEEE Transactions on Multimedia 20, 6 (2018), 1499–1511.[61] City of Lausanne. 2010. Drinking in Public Space in City of Lausanne: A Challenge for Police. http://bit.ly/28M8X8X
Accessed: July 15, 2016.[62] Vicente Ordonez and Tamara L Berg. 2014. Learning high-level judgments of urban perception. In European Conference
on Computer Vision. Springer, 494–510.[63] Stephen A Petrill, Alison Pike, Tom Price, and Robert Plomin. 2004. Chaos in the home and socioeconomic status are
associated with cognitive development in early childhood: Environmental mediators identified in a genetic design.Intelligence 32, 5 (2004), 445–460.
[64] Thanh-Trung Phan and Daniel Gatica-Perez. 2017. Healthy# fondue# dinner: analysis and inference of food anddrink consumption patterns on instagram. In Proceedings of the 16th International Conference on Mobile and UbiquitousMultimedia. ACM, 327–338.
[65] Lorenzo Porzi, Samuel Rota Bulò, Bruno Lepri, and Elisa Ricci. 2015. Predicting and understanding urban perceptionwith convolutional neural networks. In Proceedings of the 23rd ACM international conference on Multimedia. ACM,139–148.
[66] Daniele Quercia, Neil Keith O’Hare, and Henriette Cramer. 2014. Aesthetic capital: what makes London look beautiful,quiet, and happy?. In Proceedings of the 17th ACM conference on Computer supported cooperative work & social computing.ACM, 945–955.
[67] Miriam Redi, Luca Maria Aiello, Rossano Schifanella, and Daniele Quercia. 2018. The Spirit of the City: Using SocialMedia to Capture Neighborhood Ambiance. Proceedings of the ACM on Human-Computer Interaction 2, CSCW (2018),144.
[68] Miriam Redi, Daniele Quercia, Lindsay T Graham, and Samuel D Gosling. 2015. Like partying? your face says it all.predicting the ambiance of places with profile pictures. arXiv preprint arXiv:1505.07522 (2015).
[69] Harriet L Rheingold and Kaye V Cook. 1975. The contents of boys’ and girls’ rooms as an index of parents’ behavior.Child development (1975), 459–463.
[70] RonPorter. [n.d.]. An example of bedroom on Pixabay. Retrieved Aug 19, 2019 from https://pixabay.com/photos/living-room-great-room-670240/
[71] Allison Sall and Rebecca E Grinter. 2007. Let’s get physical! In, out and around the gaming circle of physical gaming athome. Computer Supported Cooperative Work (CSCW) 16, 1-2 (2007), 199–229.
[72] Darshan Santani, Joan-Isaac Biel, Florian Labhart, Jasmine Truong, Sara Landolt, Emmanuel Kuntsche, and DanielGatica-Perez. 2016. The night is young: urban crowdsourcing of nightlife patterns. In Proceedings of the 2016 ACMInternational Joint Conference on Pervasive and Ubiquitous Computing. ACM, 427–438.
[73] Darshan Santani and Daniel Gatica-Perez. 2013. Speaking swiss: languages and venues in foursquare. In Proceedings ofthe 21st ACM international conference on Multimedia. ACM, 501–504.
[74] Darshan Santani and Daniel Gatica-Perez. 2015. Loud and trendy: Crowdsourcing impressions of social ambiance inpopular indoor urban places. In Proceedings of the 23rd ACM international conference on Multimedia. ACM, 211–220.
[75] Darshan Santani, Rui Hu, and Daniel Gatica-Perez. 2016. InnerView: Learning place ambiance from social mediaimages. In Proceedings of the 2016 ACM on Multimedia Conference. ACM, 451–455.
[76] Klaus R Scherer, Marcel R Zentner, et al. 2001. Emotional effects of music: Production rules. Music and emotion: Theoryand research 361, 2001 (2001), 392.
[77] Diane J Schiano, Ame Elliott, and Victoria Bellotti. 2007. A look at Tokyo youth at leisure: Towards the design of newmedia to support leisure outings. Computer Supported Cooperative Work (CSCW) 16, 1-2 (2007), 45–73.
[78] Robert Shaw. 2015. Controlling darkness: self, dark and the domestic night. cultural geographies 22, 4 (2015), 585–600.[79] Patrick E Shrout and Joseph L Fleiss. 1979. Intraclass correlations: uses in assessing rater reliability. Psychological
bulletin 86, 2 (1979), 420.[80] Ernestasia Siahaan, Alan Hanjalic, and Judith Redi. 2016. A reliable methodology to collect ground truth data of image
aesthetic appeal. IEEE Transactions on Multimedia 18, 7 (2016), 1338–1350.[81] Tracey Skelton and Gill Valentine. 2005. Cool places. In Cool Places. Routledge, 11–42.[82] Nanette Stroebele and John M De Castro. 2004. Effect of ambience on food intake and food choice. Nutrition 20, 9
(2004), 821–838.[83] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. 2016. Rethinking the inception
architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition.2818–2826.
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
[84] Eran Toch and Inbal Levi. 2013. Locality and privacy in people-nearby applications. In Proceedings of the 2013 ACMinternational joint conference on Pervasive and ubiquitous computing. ACM, 539–548.
[85] Jasmine Truong. 2018. Attending to others: how digital technologies direct young people’s nightlife. GeographicaHelvetica 73, 2 (2018), 193–201.
[86] Jasmine Truong. 2018. Collapsing contexts: social networking technologies in young people’s nightlife. Children’sGeographies 16, 3 (2018), 266–278.
[87] Yi-Fu Tuan. 1971. Geography, phenomenology, and the study of human nature. Canadian Geographer/Le Géographecanadien 15, 3 (1971), 181–192.
[88] Ian Tucker. 2010. Everyday spaces of mental distress: the spatial habituation of home. Environment and Planning D:Society and Space 28, 3 (2010), 526–538.
[89] Gill Valentine, Sarah Holloway, Charlotte Knell, and Mark Jayne. 2008. Drinking places: Young people and cultures ofalcohol consumption in rural environments. Journal of Rural Studies 24, 1 (2008), 28–40.
[90] Ilse Van Liempt, Irina Van Aalst, and Tim Schwanen. 2015. Introduction: Geographies of the urban night.[91] viganhajdari. [n.d.]. An example of living room on Pixabay. Retrieved Aug 19, 2019 from https://pixabay.com/photos/
interior-decoration-home-design-1519595/[92] He Wang, Dimitrios Lymberopoulos, and Jie Liu. 2014. Local business ambience characterization through mobile audio
sensing. In Proceedings of the 23rd international conference on World wide web. ACM, 293–304.[93] Clifton Edwin Watts. 2004. Predicting Free Time Activity Involvement of Adolescents: The Influence of Adolescent
Motivation, Adolescent Initiative, and Perceptions of Parenting. (2004).[94] Samantha Wilkinson. 2017. Drinking in the dark: shedding light on young people’s alcohol consumption experiences.
Social & Cultural Geography 18, 6 (2017), 739–757.[95] Samantha Wilkinson and Catherine Wilkinson. 2018. Night-Life and Young People’s Atmospheric Mobilities. Mobile
Culture Studies. The Journal 3, 2017 (2018), 77–96.[96] Richmond YWong, Deirdre KMulligan, Ellen VanWyk, James Pierce, and John Chuang. 2017. Eliciting values reflections
by engaging privacy futures using design workbooks. Proceedings of the ACM on Human-Computer Interaction 1,CSCW (2017), 111.
[97] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. 2017. Places: A 10 million ImageDatabase for Scene Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017).
[98] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. 2018. Places: A 10 million imagedatabase for scene recognition. IEEE transactions on pattern analysis and machine intelligence 40, 6 (2018), 1452–1464.
[99] Jonathan Zong and J Nathan Matias. 2018. Automated Debriefing: Interface for Large-Scale Research Ethics. InCompanion of the 2018 ACM Conference on Computer Supported Cooperative Work and Social Computing. ACM, 21–24.
Received April 2019; revised June 2019; accepted August 2019
Proc. ACM Hum.-Comput. Interact., Vol. 3, No. CSCW, Article 189. Publication date: November 2019.
